Cohere Rerank 3

Rerank is a narrow product, and that narrowness is the point. It does one job: take a list of candidate documents and reorder them by how well each actually answers the query. Because the model scores query and document jointly rather than comparing two embeddings, it catches relevance that vector similarity alone misses — a fact the vendor leans on hard in pitching the economic case, since passing fewer documents into the generator cuts both token spend and latency and gives agent traces less noise to drag around. The most interesting thing on the current vendor page is Context Focus. Instead of handing back a list of document IDs with scores and making you re-chunk them, the call returns scored relevant spans alongside the document scores, and you set thresholds for how much text to keep. For teams building RAG that is a real pipeline simplification — one model call instead of a rerank plus a separate extraction step — and it is the feature most worth testing against whatever you run today. Rerank 4, per the December 11, 2025 changelog, improved accuracy and speed over earlier rerank models, covers 100+ business languages, and ranks emails, tables, JSON, and code, not just prose. That semi-structured coverage is where dense embeddings tend to fail and where a cross-encoder earns its fee. Where it will not help: Rerank reorders, it does not retrieve. If you have no first-pass stage producing candidates, there is nothing to rerank, and the vendor's own product family expects you to pair it with something like Embed. It also adds a stage of latency per query, which interactive search UIs need to measure rather than assume away. Chunking, indexing, and orchestration stay your problem. On deployment, you can call the hosted /rerank endpoint, reach it through cloud providers (Azure AI is named), or run dedicated inference in your own VPC via Model Vault — $5.00/hr or $3,250/mo for Rerank 4 Fast Medium and Pro Medium, $10.00/hr or $6,500/mo for Pro Large, billed hourly or monthly. That VPC option is the reason regulated buyers shortlist Cohere against self-hosted open-source cross-encoders: you trade engineering time for infrastructure control either way, and the right answer depends on which is genuinely scarcer for you.